Sea surface temperature (SST) prediction has received increasing attention in recent years due to its paramount importance in the various fields of oceanography. Existing studies have shown that neural networks are particularly effective in making accurate SST predictions by efficiently capturing spatiotemporal dependencies in SST data. Among various models, the ConvLSTM framework is notably prominent. This model skillfully combines convolutional neural networks (CNNs) with recurrent neural networks (RNNs), enabling it to simultaneously capture spatiotemporal dependencies within a single computational framework. To overcome the limitation that CNNs primarily capture local spatial information, in this paper we propose a novel model named DatLSTM that integrates a deformable attention transformer (DAT) module into the ConvLSTM framework, thereby enhancing its ability to process more complex spatial relationships effectively. Specifically, the DAT module adaptively focuses on salient features in space, while ConvLSTM further captures the temporal dependencies of spatial correlations in the SST data. In this way, DatLSTM can adaptively capture complex spatiotemporal dependencies between the preceding and current states within ConvLSTM. To evaluate the performance of the DatLSTM model, we conducted short-term SST forecasts in the Bohai Sea region with forecast lead times ranging from 1 to 10 days and compared its efficacy against several benchmark models, including ConvLSTM, PredRNN, TCTN, and SwinLSTM. Our experimental results show that the proposed model outperforms all of these models in terms of multiple evaluation metrics short-term SST prediction. The proposed model offers a new predictive learning method for improving the accuracy of spatiotemporal predictions in various domains, including meteorology, oceanography, and climate science.

Accurate prediction of sea surface temperature (SST) is crucial in the field of oceanography, as it has a significant impact on various physical, chemical and biological processes in the marine environment. In this study, we propose a physics-guided attention-based neural network (PANN) to address the spatiotemporal SST prediction problem. The PANN model incorporates data-driven spatiotemporal convolution operations and the underlying physical dynamics of SSTs using a cross-attention mechanism. First, we construct a spatiotemporal convolution module (SCM) using ConvLSTM to capture the spatial and temporal correlations present in the time series of the SST data. We then introduce a physical constraints module (PCM) to mimic the transport dynamics in fluids based on data assimilation techniques used to solve partial differential equations (PDEs). Consequently, we employ an attention fusion module (AFM) to effectively combine the data-driven and PDE-constrained predictions obtained from the SCM and PCM, aiming at enhancing the accuracy of the predictions. To evaluate the performance of the proposed model, we conduct short-term SST forecasts in the East China Sea (ECS) with forecast lead times ranging from 1 to 10 days, by comparing it with several state-of-the-art models, including ConvLSTM, PredRNN, TCTN, ConvGRU, and SwinLSTM. The experimental results demonstrate that our proposed model outperforms these models in terms of multiple evaluation metrics for short-term predictions.

Sea surface temperature (SST) is a fundamental parameter in the field of oceanography as it significantly influences various physical, chemical, and biological processes within the marine environment. In this study, we propose an attention-based context fusion network (ACFN) model for short-term prediction of SST based on the operational SST and sea ice analysis (OSTIA) data. The ACFN model integrates an attention-based context fusion block into the convolutional long short-term memory (ConvLSTM) model. Specifically, the attention mechanism generates attention maps sequentially across both the channel and spatial dimensions, allowing for the detailed exploration of intricate spatiotemporal correlations between the previous context state and the current input state in ConvLSTM. To assess the performance of the ACFN model, we apply it to predict SST in the Bohai Sea with lead times ranging from one to ten days. The results demonstrate that our proposed model outperforms several state-of-the-art models, i.e., ConvLSTM, predictive RNN (PredRNN), SimVP, and motion details RNN (MoDeRNN), in terms of mean absolute error (MAE) and coefficient of determination ( $R^{2}$ ). In particular, our analysis reveals that the prediction errors are relatively higher in the coastal areas compared to those in the central Bohai Sea.

Traffic density is a crucial indicator for evaluating the level of service, as it directly reflects the degree of road congestion and driving comfort. However, accurately predicting real-time traffic density has been a significant challenge in Intelligent Transportation Systems (ITS) due to the nonlinear and spatial-temporal dynamic complexity of traffic density. In this paper, we propose a novel Model-enhanced Spatial-Temporal Attention Network (MSTAN), which constructs a spatial-temporal traffic kernel density model using the Kernel Density Estimation (KDE) method to process the spatiotemporal data and calculate the probabilities of various spatiotemporal events. These probabilities are input into the attention mechanism, enabling the model to recognize the inherent connection between dynamic and distant events. Through this fusion, the network can deeply learn and analyze the spatial-temporal properties of traffic features. Furthermore, this paper utilizes the attention mechanism to dynamically model spatial-temporal dependencies, capturing real-time traffic conditions and density, and constructs a spatial-temporal attention module for learning. To validate the performance of the proposed MSTAN model, experiments are conducted on two public datasets of California highways (PeMS04 and PeMS08). The experimental results demonstrate that the MSTAN model outperforms existing state-of-the-art baseline models in terms of prediction accuracy, thus proving the effectiveness of the model both theoretically and practically.

Numerical weather prediction of sea surface temperature (SST) is crucial for regional operational forecasts. Deep learning offers an alternative approach to traditional numerical general circulation models for numerical weather prediction. In our previous work, we developed a sophisticated deep learning model known as the Attention-based Context Fusion Network (ACFN). This model integrates an attention mechanism with a convolutional neural network framework. In this study, we applied the ACFN model to the South China Sea to evaluate its performance in predicting SST. The results indicate that for a 1-day lead time, the ACFN model achieves a Mean Absolute Error of 0.215 °C and a coefficient of determination ($R^2$) of 0.972. In addition, in situ buoy data were utilized to validate the forecast results. The Mean Absolute Error for forecasts using these data increased to 0.500 °C for a 1-day lead time, with a corresponding $R^2$ of 0.590. Comparative analyses show that the ACFN model surpasses traditional models such as ConvLSTM and PredRNN in terms of accuracy and reliability.

Accurate prediction of sea surface temperature (SST) and sea level anomaly (SLA) is essential for numerical weather prediction, climate monitoring, and ocean state estimating, as both variables are key indicators of the thermal and dynamic conditions of the upper ocean. However, effectively capturing the complex dynamic coupling between SST and SLA, along with their spatiotemporal dependencies and mutual influences, remains a significant challenge. This study proposes a unified deep learning framework, termed CA-convolutional long short-term memory (ConvLSTM), which jointly predicts SST and SLA by integrating a cross-attention mechanism with a ConvLSTM architecture. The cross-attention module enables dynamic feature interaction between SST and SLA, allowing us to effectively capture complex spatiotemporal dependencies. A key advantage of cross-attention-enhanced ConvLSTM (CA-ConvLSTM) lies in its ability to simultaneously learn from SST and SLA through a unified training process, thus improving the prediction accuracy of both variables. Numerical experiments conducted in the South China Sea demonstrate that the proposed CA-ConvLSTM consistently outperforms several state-of-the-art baseline models, including ConvLSTM, PredRNN, TCTN, SwinLSTM, and UniOcean, in the joint prediction of SST and SLA. These results underscore the effectiveness of incorporating multivariate oceanic information through attention-based mechanisms and highlight the potential of such frameworks for advancing spatiotemporal marine prediction capabilities.